To Do List

[luckynozomi]

• [ ] Add support for local response layers
• [x] Optimize for GPU performance
• [x] Export and Import training results
• [ ] Parameter Selection in CNN
• [ ] Write logistic regression as a subclass of FCLayer
• [ ] Make the code more easy to use, no need to modify the code to reconstruct the layer structure
• [x] Data augmentation in training
• [ ] Force every kernel to learn independent feature
• [ ] Selection of activation function and loss function
• [x] Subset the data to get a dataset to experiment on
• [ ] Learn soft weight sharing and find paper on the performance
• [ ] Learn early stopping and understand more complicated gradient descent algorithms

[luckynozomi]

• Data augmentation in training is implemented.
  • Reflection
  • Random shifting

[luckynozomi]

• GPU performance optimized for cuDNN
  • use FFT for faster convolution
  • code optimized when needs padding
• Export and Import training results implemented
• No need to subset the data to experiment on